The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill iterates through and processes content from existing memory files to resolve contradictions and remove stale facts. This ingestion of data from previous sessions creates an attack surface where malicious instructions could be embedded in memory files to influence the agent's consolidation logic.
Ingestion points: Reads content from all Markdown files within discovered memory directories, including platform-specific paths like ~/.claude/projects/.
Boundary markers: The skill does not employ specific delimiters or "ignore embedded instructions" warnings when analyzing the content of memory files.
Capability inventory: The agent has the capability to read, write, and delete files on the local system, and utilizes tools like git, grep, and glob for verification.
Sanitization: No explicit content sanitization or validation is described; the skill relies on the agent's internal reasoning to differentiate between data and instructions.
[COMMAND_EXECUTION]: The skill logic relies on the execution of repository-level tools including git (to verify modification dates), grep (to check for references in code), and glob (to verify file existence). This provides the agent with broad read access to the project structure and metadata to perform its maintenance tasks.

dream